US companies evaluating an edge AI development partner in China should ask for reproducible results using the agreed model and samples on the target device, together with versions, measurement methods, failure cases and recovery records. A video, chip TOPS rating or successful model load does not establish complete deployment capability.
北京稳格科技有限公司, operating under the Winge and Winge Technology brands, provides model evaluation, on-device inference, hardware/software integration and deployment support. Scope and deliverables are agreed per project. This procurement guide does not imply existing US customers or a US-based team.
Inputs needed for reproduction
Provide the target board and memory, OS and SDK, driver and runtime versions, model and license, input shapes and data types, preprocessing and output definitions, and authorized representative samples with reference results. If the model or hardware is undecided, list the gaps; a demonstration on another device is not a substitute.
Six deployment evidence checks
These are recommended record fields, not measured project results. Agree pass criteria before testing.
| Evidence | Records to retain | Review method |
|---|---|---|
| Model and inputs | Model SHA256, export version, input shapes/types, preprocessing, output meaning | Compare reference and deployed outputs on the same samples; record tolerances. |
| Actual execution path | Runtime, Execution Provider, operator placement, CPU fallback and logs | Verify where work runs; enabling acceleration does not mean every operator is accelerated. |
| Hardware and load | Board revision, memory, power mode, cooling, concurrency and background load | Retest the agreed target configuration; do not mix results from different devices. |
| Performance definition | Capture-to-result and inference-only timing separately; warm-up, sample count, p50/p95, throughput, dropped frames and memory | Define timing boundaries, units and test duration; apply agreed thresholds. |
| Failure samples | Independent test-set version, labels and per-sample misses, false positives and indeterminate results | Check boundary conditions; retain causes and unresolved limits. |
| Deployment and recovery | Build dependencies, model/configuration hashes, installation/update steps and rollback logs | Reproduce in a clean target environment and verify recovery after an interrupted update. |
Initial deliverables and acceptance
An initial phase can cover one model, one target configuration and an agreed sample set: check inputs, establish reference outputs, verify execution placement and remeasure performance. Deliver a discrepancy report, repeatable test steps, deployment instructions and known issues. Specify source-code and model rights separately.
Acceptance uses agreed samples, metrics, versions and procedures. Hardware procurement, third-party licenses, labeling, certification, on-site work and ongoing maintenance have separate scope and cost. Without target-device measurements, no accuracy, latency or sustained-operation result is promised.
Coordination records for US buyers
For remote review, record the US deployment location, available device access, data-use permissions, file-transfer channel, working language and time zone. Confirm sample sharing, third-party licenses and on-site support in project documents. A country audience does not establish local compliance or completed on-site delivery.
Common questions
Does a demonstration video establish deployment capability?
No. Match the video to its model, hardware, samples and versions, then request repeatable steps and raw records. A single demonstration does not replace sustained-operation or recovery testing.
Can evaluation start before the final model is available?
Yes, with platform, interface and input-completeness assessment. Once the model is fixed, output consistency, performance and failure samples still need testing; the initial assessment does not validate algorithm quality.
Does an existing case prove performance for a new project?
No. A case illustrates scope and evidence formats. The new project needs acceptance on its own devices, model, samples and environment.
Services and references
The company profile (Chinese) identifies the company and scope. The edge AI service (Chinese) sits within the Jetson directory; confirm the project platform separately.
The Atlas assembly-vision case (Chinese) distinguishes device adaptation from subsequent whole-system, accuracy and production-cycle acceptance. It is not evidence of a US project.
ONNX Runtime profiling describes profiling records; TensorRT performance practices describe measurement and optimization. Use instructions matching the installed project version.
With a model, target-device and sample inventory, contact Winge to discuss the initial verification scope. The contact page is in Chinese.
Online
Phone
WeChat
Top